Trustworthy Semi-Supervised Machine Learning

Semi-supervised learning has become an important machine learning paradigm for reducing the dependence on costly labelled data by exploiting abundant unlabelled data. However, in real-world applications, especially in high-stakes domains such as biometrics, healthcare, finance, security, and autonomous systems, model accuracy alone is not sufficient. A semi-supervised learning system should not only perform well, but also be reliable, safe, fair, understandable, and privacy-preserving.In this Research Topic, trustworthiness refers to the ability of a semi-supervised machine learning system to produce dependable and responsible outcomes under realistic conditions. Specifically, it includes robustness against noisy inputs, adversarial perturbations, distribution shifts, and unreliable pseudo-labels; fairness across different demographic or user groups; explainability and interpretability that help users understand model decisions; privacy protection for sensitive training data; and well-calibrated confidence or uncertainty estimates that indicate when predictions can or cannot be trusted.This Research Topic aims to advance trustworthy semi-supervised machine learning by encouraging new theories, algorithms, evaluation protocols, and applications that explicitly address these dimensions of trustworthiness. Particular attention will be given to uncertainty quantification, model calibration, and the risks of pseudo-labelling, including noisy pseudo-labels, confirmation bias, error propagation, and overconfident predictions. The distinctive contribution of this Topic is to provide a focused venue for research that moves semi-supervised learning beyond label efficiency toward safe, fair, transparent, and dependable AI deployment.Traditional semi-supervised learning has mainly focused on improving predictive performance by using unlabelled data through pseudo-labelling, self-training, consistency regularization, graph-based learning, generative modelling, and representation learning. These approaches are effective when labelled data are limited, but they may introduce important trustworthiness risks. For example, incorrect pseudo-labels can be repeatedly reinforced during training, leading to confirmation bias and accumulated errors. Poorly calibrated confidence scores may cause a model to treat uncertain predictions as reliable labels. Unlabelled data may also contain distribution shifts, class imbalance, demographic bias, or out-of-distribution samples.Meanwhile, recent advances in multimodal learning, foundation models, federated learning, continual learning, and biometric applications have created new opportunities for semi-supervised learning. Foundation models and multimodal large models can support semi-supervised adaptation with limited labelled data, but their predictions and pseudo-labels may be biased or overconfident. Federated learning can enable semi-supervised training across decentralized data while protecting privacy, but it also faces data heterogeneity and reliability challenges. Continual learning allows models to adapt over time, but may accumulate pseudo-label errors. Biometrics, such as face, fingerprint, iris, gait, voice, and multimodal identity recognition, is an important application context because labelled data are sensitive and errors may affect privacy, fairness, and security.This Research Topic welcomes original research articles, reviews, perspectives, and methodology papers that advance trustworthy semi-supervised machine learning. All submissions must demonstrate a substantive connection to semi-supervised learning by clearly describing the roles of labelled and unlabelled data, how unlabelled data are used, and how the work contributes to semi-supervised methods, theory, evaluation, or applications. Papers focused only on federated, multimodal, continual, biometric, foundation-model-based, or general trustworthy AI methods without a central semi-supervised component will be outside the scope.Topics of interest include trustworthy pseudo-labelling, self-training, consistency regularization, graph-based learning, generative modelling, and representation learning; robustness, fairness, explainability, privacy, uncertainty quantification, calibration, and pseudo-label risk mitigation; semi-supervised learning for biometrics, healthcare, autonomous driving, and security-critical applications; federated, continual, multimodal, and foundation-model-based semi-supervised learning; and benchmarks, protocols, metrics, reproducibility studies, and deployment frameworks for trustworthy semi-supervised settings. Topic editors: Chenglong Zhang, Jinxing Li Source: https://www.frontiersin.org/research-topics/83434/trustworthy-semi-supervised-machine-learning